[Yunnglin]

The CC-OCR-V2 documentation still reports 7,091 loaded samples and attributes two missing images to the official dataset. Historical ModelScope mirror revisions reverted those images; both are present in the current mirror and official Hugging Face dataset. A fresh snapshot now loads all 7,093 samples with the existing adapter.

This updates the source description, regenerates full-dataset statistics and bilingual benchmark pages, and adds regression coverage for the two samples being included once their images are restored. The grounding description now explicitly explains that absolute pixel predictions are interpreted as 0-1000 coordinates and scaled to pixels. Fixed usage hints are localized as Chinese comments while code blocks and string/data literals remain protected. The shared benchmark-documentation translator now uses qwen3.8-flash. Scoring and loading semantics are unchanged.

Validation:

• Full HF/ModelScope sample-file manifests agree; all 7,093 question/answer/image sets are complete.
• Actual adapter scan: 7,093 samples across 16 subsets, including both reported IDs.
• Both real restored samples pass gold-answer replay; a two-sample qwen3-vl-plus evaluation produces predictions, successful reviews, and a two-sample report.
• 16 benchmark tests plus 4 translation tests, CI lite, and make lint pass.
• Metadata and bilingual documentation regenerated through make docs-pipeline BENCHMARK=cc_ocr_v2 FORCE=1 and make docs-generate.

Maintainer clarification: Eioss/CC-OCR-V2#3 (comment)

Historical mirror evidence: revision aca514a4cc7b662b14b24d9c755ab770b8632e2e (2026-08-10) has 149 images in doc_parsing_doc_doc_scan_150, excluding 4a74943cc5a4.jpg. Current revision 5acd45e358b64ca426f079e2bb52d56a4853bbb2 has all 150. Commit history also records the corresponding image revert for 0913e4fab7c9.jpg.

[Copilot]

🟢 Approval recommended

No active review issues were identified in the CC-OCR-V2 documentation, metadata, and regression-test changes.

No open issues 🎉

What changed in this PR

Updates EvalScope’s CC-OCR-V2 documentation to reflect all 7,093 samples now available in the complete dataset, without changing loading or scoring semantics.

Changes:

• Corrects the dataset description and adds guidance for diagnosing incomplete snapshots.
• Refreshes generated metadata and bilingual pages to report 7,093 samples and 7,118 images.
• Adds regression tests confirming both affected samples load once their missing images are restored.

File	Description
tests/​benchmark/​test_cc_ocr_v2.py	Adds restored-image regression coverage for both sample IDs.
evalscope/​benchmarks/​cc_ocr_v2/​cc_ocr_v2_adapter.py	Updates description text and incomplete-snapshot guidance only.
evalscope/​benchmarks/​_meta/​cc_ocr_v2.json	Refreshes generated statistics and bilingual documentation content.
docs/​zh/​benchmarks/​cc_ocr_v2.md	Regenerates the Chinese page with complete dataset counts.
docs/​en/​benchmarks/​cc_ocr_v2.md	Regenerates the English page with complete dataset counts.

_{Review effort: Balanced}